fix: use OpenAI chat-completion field names in /chat/completions usage

[chilang]

Summary

UsageStats (used for /v1/chat/completions responses) inherits from OpenAIUsage, which models the OpenAI Responses API (/v1/responses) — that spec uses input_tokens / output_tokens. But /v1/chat/completions is a different spec: the usage object must contain prompt_tokens, completion_tokens, and total_tokens.

The net effect: any OpenAI-compatible client hitting mlx-vlm's /v1/chat/completions fails to read the usage payload because the field names don't match. I hit this reproducing Gemma 4 benchmarks with llama-benchy, which errors out during warmup with Warmup failed: 'prompt_tokens'.

Fix

• Stop inheriting UsageStats from OpenAIUsage. Keep OpenAIUsage as-is for the Responses API (/v1/responses) where the spec is correct.
• Declare the chat-completion fields directly on UsageStats:
  • prompt_tokens: int
  • completion_tokens: int
  • total_tokens: int
  • (existing prompt_tps / generation_tps / peak_memory extras preserved)
• Update the two call sites in chat_completions_endpoint (streaming SSE chunk + non-streaming final response) to build UsageStats with the new field names.
• /v1/responses is untouched — it keeps using OpenAIUsage with input_tokens / output_tokens per OpenAI's Responses API spec.

Test plan

• Added regression test test_chat_completions_response_uses_openai_usage_field_names that mocks generate() and asserts the JSON response body contains usage.prompt_tokens, usage.completion_tokens, usage.total_tokens, and does not contain the Responses-API field names.
• python -m pytest mlx_vlm/tests/test_server.py — 10 passed.
• Manually verified with curl against /v1/chat/completions on a locally running server — the usage object now matches the OpenAI Chat Completions spec.
• Verified with llama-benchy against mlx-community/gemma-4-E4B-it-4bit and mlx-community/gemma-4-26b-a4b-it-4bit — warmup now succeeds and benchmarks complete.
• black --check and isort --profile=black --check pass on changed files.

References

• OpenAI Chat Completions: https://platform.openai.com/docs/api-reference/chat/object — usage object fields prompt_tokens / completion_tokens / total_tokens
• OpenAI Responses: https://platform.openai.com/docs/api-reference/responses/object — usage object fields input_tokens / output_tokens